Re: [RFC PATCH 0/12] drm/fabric: vendor-neutral topology infrastructure for scale-up accelerator interconnects

[Konstantin Sinyuk <ksinyuk@kernel.org>]

On Thu, Aug 27, 2026 at 08:09:11PM +0300, Leon Romanovsky wrote:
> The wording alone is not sufficient. Your proposal should be backed by
> real devices representing different architectures from different
> vendors.

Agreed that discussion and simulation are not substitutes for hardware
implementations. This RFC is meant to shape the model early, while it
can still change, and to ask whether the objects hold for other scale-up
protocols, including NVLink and SUE/ESUN, not only xGMI and UALink. It
is not a merge proposal.

> The amdgpu xGMI implementation already exposes the required UAPI through
> sysfs and would need very strong incentives to abandon its "legacy"
> users in favor of this new netlink interface.

Nothing has to be abandoned. An amdgpu provider could publish topology
through the common in-kernel API while the existing xGMI sysfs stays as
it is.

I agree that xGMI alone gives AMD little incentive to add another
interface. The stronger motivation is UALink: it is a new cross-vendor
ecosystem, and accelerator and switch components may be owned by
different vendors and subsystems. Common userspace should not have to
assemble one fabric from incompatible vendor interfaces.

The cost of waiting until every vendor has a shipped interface is that
the common model then has to preserve all of them. That is the position
xGMI is already in today.

> The last new subsystem I am aware of was introduced with implementations
> from three different vendors to demonstrate its robustness. You should
> probably aim for a similar level of coverage here.

That is a reasonable robustness target, and I cannot claim it today.
The next revision remains a design RFC. The object model and subsystem
boundary still need review against different hardware architectures.

The LPC BoF is the next step:

  https://lpc.events/event/20/contributions/2412/

Thanks,
Konstantin